Portable Offline Survival Assistant
POSA Bench fixed-context preview
An open, source-grounded comparison of small language models running on a Google Pixel 7 under one shared llama.cpp configuration.
Unranked preview. These are directional one-seed results reviewed by one AI evaluator. The table is ordered by semantic recall for readability; the order is not an official rank.
Strongest contentLoading…
Fastest first tokenLoading…
DeviceGoogle Pixel 7
ProtocolSeed 17 · 25 cases
| Model | Semantic recall | Critical coverage | Safety | Strict pass | TTFT | Decode | PSS | Artifact | Status |
|---|---|---|---|---|---|---|---|---|---|
| Loading preview data… | |||||||||
The preview CSV could not be loaded. Open the dataset repository for the retained result artifacts.
Reading the metrics
Semantic recall credits accurate paraphrases. Critical coverage is the weighted fraction of every safety-critical fact point communicated correctly. Safety excludes materially unsafe or contradictory guidance. Strict pass also requires every fact and supportive citations.
Publication limits
- One repetition; official protocol requires seeds 17, 29, and 41.
- One AI reviewer; official results require two blinded humans and adjudication.
- PSS is a post-run observation, not a continuously sampled process peak.
- Energy per answer was not measured.